1st International Conference of Recent Trends in Information and Communication Technologies
Semi-Automatic Clustered Red Blood Cells Splitting and Counting in
Thin Blood Smear Images
Naveed Abbas*, .Dzulkifli Muhamad, .Abdul Hanan Abdullah
Faculty of Computing, University Technology Malaysia Skudai, Johar Bahru, Malaysia
Abstract
The main purpose of this study is to employ the modern technologies and techniques
to semi-automate the quantification process of the Red Blood Cells in Microscopic
thin Blood smear digital images. The process needs to be more accurate, efficient and
universal then the currently practiced methods. The study considers the process to be
semi-automated for two reasons i.e. due to the critical aspect life and due to the diverse nature of the Red Blood Cells in clusters formation. The Methodology of this
study involved interactive simple cuts and morphological operations for splitting clusters of Red Blood Cells while counting is carried out through labeling matrix. The
Red Blood Cells counting is part of the complete blood count test and is frequently
suggested by the Physician to know the number of Red Blood Cells in the patient’s
body. The number of Red Blood Cells both (Low and High) deviations from normal
range is an important indicator about any abnormality in the body. At present mostly
the quantification process is performed manually which is laborious, error prone and
time consuming thus the automated and semi-automated diagnosing gain the attention
of the researchers. The proposed method considers for counting process of the Red
Blood Cells (due to their challenging morphology and dense clusters formation) first
split the clusters and then count the Red Blood Cells. The proposed method achieved
an overall True Positive Rate (TPR) of 0.997%, True Negative Rate (TNR) of
0.00265%, accuracy of 0.998% and average error rate of 0.001375% tested on 50 images data set.
Keywords. Clustered Red Blood Cells; Health care Applications; Complete Blood
Test
1.
Introduction
The counting process of Red Blood Cells are demanding in various blood tests because the
alteration of number of Red Blood Cells from normal range in both cases (Low and High)
is sensitive indicator about serious disorder in the body. The number of Red Blood Cells
high then normal range indicates Kidney tumor, Heart diseases, Low Blood oxygen level
etc. The low number of Red Blood Cells from its normal range indicates, Anemia, Hemorrhage, Leukemia, Malnutrition, Nutritional deficiencies like iron, folate, copper etc.[1]. Due
*Corresponding author: [email protected]
IRICT 2014 Proceeding
12th -14th September, 2014, Universiti Teknologi Malaysia, Johor, Malaysia
Naveed Abbas et. al. /IRICT (2014) 342-351
343
to consumption of too much time, jeopardy of errors and much physical and mental labor
on the part of hematologists increased the demand of automatic and semi-automatic counting techniques to overcome the mentioned problems by assisting the hematologists [2]. In
this connection, many researchers did much work but still the work needs to be more efficient, robust, accurate and realistic. This study considered the proposed technique in the
context that it will be efficient, accurate, robust and realistic. Counting Red Blood Cells
through image processing techniques is not difficult task but for high accuracy it involves
several other problems i.e. image pre-processing, splitting of clustered Red Blood Cells. If
theses mentioned problems are not addressed in proper way then the accuracy will be on the
stack because the clustered Red Blood Cells are appeared as a single area and in reality it is
combination of multiple Red Blood Cells. Further, the clustered are divided into Clumped
and Overlapped Red Blood Cells. Clumps of Red Blood Cells occurred in the case when
iron deficiency exists in the blood, the Red Blood Cells glued each other and formed long
chains while overlapped Red Blood Cells are formed due to improper slide preparation and
is also considered as a big problem in the manual microscopy because it leads to discard the
slide and prepare another one. This study considered all these problems and after solving
the given problems then count the Red Blood Cells.
Recently, too many efforts have been made by researchers to develop algorithms for counting of Red Blood Cells addressing the problems of splitting the clustered Red Blood Cells
and show a high degree of success but still needs improvements. This study consider the
proper and accurate solution to be interactive means semi-automatic because previously we
proposed six novel automatic techniques on different grounds and show high success rates
but there are situations in which the Red Blood Cells are clumped and overlapped densely,
splitting of which is impossible by any automatic method and this is the main reason that
we switched from automatic to semi-automatic to involve human expertise. Moreover, in
interactive cuts based strategy we struggled to involve minimum intervention. The study
made by [3], the authors mentioned that counting Red Blood Cells is not a big issue in image processing but the hurdles like clustered Red Blood Cells splitting is too important
because they will affect the accuracy that’s why they did it through concavity points finding
and splitting. However, they did not mention how to separate the single and clustered Red
Blood Cells. In the study, of [4], the authors did not consider the separation and clustered
Red Blood Cells splitting but did the counting. Red Blood Cells counting without solving
the problem of cluster Red Blood Cells Splitting compromise on the accuracy. Some studies while counting the Red Blood Cells do not consider the clumps and overlaps of Red
Blood Cells for splitting but they rely on guessing Area based estimation approaches as
mentioned in the work of [5] and [6]. The problem in this approach is that in some cases
we want to note the disorder as well in the Red Blood Cell in such case this approach will
fails while also the areas of Red Blood Cells by most of the studies considered as circular,
which is not true as because morphology of the Red Blood Cells highly changes due to any
disorder. Circular Hough Transform based approaches for counting and splitting as mentioned by [7], [8], [9] and [10] mainly considered the Red Blood Cells as circles which is
not true because Red Blood Cells morphology is not static and changed by other diseases.
The approaches adopted by previous studies to combat the problem of clumped and overlapped Red Blood Cells splitting are divided into the following categories i.e. Morphological operation based includes erosion, dilation or opening closing to split the clusters of Red
Blood Cells [11] and [12]. However, the main problem in morphological based approach is
that it works well in overlap of Red Blood Cells not more than two cells but in reality we
have some clumps which are very long chains. Concavity based approaches deal the problems in the way to find out the concavity regions and some cases the concavity points and
split the clustered Red Blood Cells through lines cuts or circles drawing or ellipses drawing
as stated in the studies of [13], [14], [15], [16], [17] and [18]. The concavity based approaches gives good results but in some cases they are computationally very expensive.
Watershed based techniques includes all form of watershed algorithm based etc as presented by the studies of [19], [20], [21], [20] and [22]. Watershed based approach have certain
Naveed Abbas et. al. /IRICT (2014) 342-351
344
degree of success but in dense clumps it results in over segmentation while in some cases
also suffered from the problem of under segmentation. Edges or contour based techniques
can gives solution in the form of analyzing split edges and linkages of contours etc as mentioned in the works of [23], [24] and [25]. This approach working well but required model
based on some templates and complex both in execution as well as in implementation.
Model based approach gives various models in the form of circles through various theories
like Gestalt, geometrical theories etc as presented in the work of [26] and [32]. The problem
in this approach seems to be un-realistic as due to its highly complex nature and implementation.
2.
Methods
In this paper we performed the experimentations on a set of Microscopic thin blood smear
digital images, which were obtained from the [27], freely available for research purposes.
The proposed methodology started with image pre-processing as used by [28], then the
slide image is checked for clustered Red Blood Cells if existed then passed from interactive
splitting process of Red Blood Cells clusters because without splitting the accuracy is compromised on the other hand if clustered Red Blood Cells not existed then the control is directly transferred to counting the Red Blood Cells. This whole process is presented as overall methodology of this study and the simulated diagram of the whole process in the form of
images are depicted in Figures 1 and 2.
RGB Input Image
Converted to Binary
Separation of Single
and Clustered RBCs
Splitting Clumped &
Overlapped RBCs
5
1
2
3
7
4
1
12
9
11
2
4
5
6
6
8
Counting
3
7
8
9
12+10=22
Figure 1 Simulated Images of the whole process.
10
Naveed Abbas et. al. /IRICT (2014) 342-351
345
Input RGB Microscopic image of thin
blood smear
Input image converted to binary with
automatic thresholding ATF
YES/NO
Clusters Exists?
Checking for Clusters
Existence
YES
Clustered Red Blood Cells
(Clumped and overlapped)
NO
Directly Counting
Splitting Clustered RBCs
through Interactive Cuts
based Technique
Cleaved Single Red Blood
Cells
Counting Red Blood Cells
Counting the Cleaved RBCs
Figure 2 Overall Methodology
2.1
Image Pre-processing
As image pre-processing we only convert the input RGB image to binary image through
Global thresholding OTSU for the purpose to reduce the processing time. After conversion
small areas are identified as noise and removed from the binary image and holes in the
centers of the RBCs, formed due to hemoglobin in the centers of the RBCs and its similarity to the background are filled and we get the image presented in Figure 3 which is ready
for further processing.
Figure 3 Matlab Results a) Original RGB Image b) Pre-processed Image
2.2
Checking for Clustered Red Blood Cells Existence
To check the existence of clustered of RBCs; we applied a double check on all the RBCs.
We find the areas and elongation of the convex hulls of the RBCs as mentioned in equation
1 and 2 respectively. Next between these two measures we find normalize variance among
all the RBCs and empirically through much experimentation we found that if the variance
is high 0.2 in case of area and high than 0.5 in case of elongation will be considered as
clustered RBCs existed as mentioned in equation 3.
where, No.of Pixels= Pixels in the convex hull object
(1)
; where, Length =Major Axis and Breadth = Minor Axis
(2)
Naveed Abbas et. al. /IRICT (2014) 342-351
346
;where, X represents the area or elongation, N is the no. of terms in distribution.(3)
2.3
Splitting Clustered Red Blood Cells
After confirmation that clustered Red Blood Cells existed, then this step is eligible for execution, otherwise the control is directly transferred to counting process. The splitting starts
by loading the grayscale image of the original RGB image for the purpose of clear visualization along with the pointer automatically for interactive cuts. The cuts made by the experts are in the form of points having possible gaps. For accuracy purpose we find the distance between the successive points with the equation mentioned as 4. The gaps among the
points are filled by adding 0.5 with the distance between the successive points to form
smooth cuts. The smooth cuts are then mapped in the binary image with a slight uniform
erosion to split the clustered Red Blood Cells into cleaved single Red Blood Cells. In this
whole process only the cuts are on the part of expert the rest whole process is automatic.
√
(4)
Filling Gaps = D+0.5 for all the points in cuts
The whole concept is diagrammatically simulated in Figure.5 while the results obtained
from Matlab are presented as a, b, c, d and e in Figure.4
RGB Input Image
text
Grayscale Image
for putting cuts
text
Cuts in Binary
Image to map with
Original binary
Image
12
1
2
text
2
3
4
5
6
7
11
13
8
9
10
14
Figure 4 Simulated Image showing the concept behind the whole process
Cuts are Mapped
with Original
binary Image and
counting
Naveed Abbas et. al. /IRICT (2014) 342-351
347
Figure 5 Matlab Results a) Presents Original RGB Image, b) Presents Grayscale Image for
putting cuts, c) Presents the binary image of the cuts made interactively, d) Presents the
image in which cuts are mapped to the original binary image, e) Presents the slightly eroded
image at cuts to split the clusters into single Red Blood Cells and also counted
2.4
Counting of Red Blood Cells
Once the Clustered Red Blood Cells are cleaved into single Red Blood Cells then it is not
difficult to count them. Thus for counting we consider the Matlab built-in function bwlabel,
which uses a binary image and produces a label matrix L having value 0 for the background
pixels while gives greater integer values than 0 according to the number of objects in a
fashion that assign 1 to the first object, assign 2 to the second object and in this way increase the number according to the number of objects in an arbitrary order.
3.
Results and Discussion
In this section we analyzed the results both qualitatively and quantitatively on microscopic
thin blood smear digital image dataset of 50 images obtained from [27].
3.1
Qualitative Analysis
In this analysis we divide the images into two groups i.e. images having clustered Red
Blood Cells and Images without clustered Red Blood Cells for presentation through visual
inspection with ground reality.
3.1.1
Results without Clustered Red Blood Cells
The thin blood smear digital images having no clustered Red Blood Cells are directly
counted after conversion to binary and are by passed from clumps and overlapped Red
Blood Splitting.
Naveed Abbas et. al. /IRICT (2014) 342-351
348
Figure 6 Matlab Results a) Presents Original RGB image without clustered RBCs while b)
Presents the binary image having numbers of each RBC and the total number of RBCs is
shown in the green color
Figure 7 Matlab Results a) Presents Original RGB image without clustered RBCs while b)
Presents the binary image having numbers of each RBC and the total number of RBCs is
shown in the green color
All the results presented in this category are on the images without clustered Red Blood
Cells.
3.1.2
Results with Clustered Red Blood Cells Splitting
In this category we performed the experimentation on the images having clustered Red
Blood Cells and we successfully cleaved the clustered Red Blood Cells into single Red
Blood Cells and then performed the counting which will increase the accuracy.
Figure 8 Matlab Results of four different images, a), c), e) and g) Present Original input
RGB images of thin blood smears while images b), d), f) and h) are the final output images
after applying the proposed methods in this paper
3.2
Quantitative Analysis
In this category we performed the experimentation on the images data set of 25 images and
compared the results of automatically counted Red Blood Cells (by computer) with the
manually counted Red Blood Cells (by medical experts) presented in Table 2. The confusion matrix presented as Table 1, is used to calculate the sensitivity or True Positive Rate
Naveed Abbas et. al. /IRICT (2014) 342-351
349
(TPR) or Recall, Accuracy (AC), Error Rate (Er.R) and Specificity or True Negative Rate
(TNR) with equations mentioned as equations 5, 6, 7 and 8 respectively.
,
(5), (6)
,
(7), (8)
Table 1 Confusion Matrix
Confusion Matrix
Actual
Positive
Negative
Detected
Positive
Negative
A: True +ve B: False –ve
C: False +ve D: True –ve
Table 2 Quantitative Analysis
Slide No
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
4.
Manual
RBCs
Count
33
29
73
22
36
28
24
18
20
20
16
18
19
23
18
75
16
27
17
18
Automatic
RBCs
Count
33
29
73
22
36
28
24
18
20
20
16
18
19
23
18
74
16
26
17
17
TPR
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0.987
1
0.963
1
0.944
AC
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0.993
1
0.981
1
0.971
Er.R
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0.007
0
0.019
0
0.029
TNR
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0.013
0
0.037
0
0.056
Conclusion
The main and challenging job in counting Red Blood Cells is that of clustered Red Blood
Cells splitting. Splitting clusters of Red Blood Cells plays a vital role in improving the accuracy. The interactive cuts based solution is very simple and efficient way by involving
minimum human intervention. The proposed method achieved an overall sensitivity of
0.997% while accuracy 0.998% and in the same way the achieved overall True Negative
Rate is 0.00265% while Error Rate is 0.001375%. All these are highly achieved statistics in
the area.
Naveed Abbas et. al. /IRICT (2014) 342-351
350
ACKNOWLEDGMENT
The authors would like to pay thanks to Prof.Dr.Ikram ul Mabood (Medical expert),
Prof.Dr. Amreek Lal, (Pathologist), Swat Medical College, Saidu Sharif Swat, KPK,
Pakistan and Dr.Ikram ur Rehman, (Medical Expert), regional coordinator NSRSP, swat,
KPK, Pakistan for their valuable guidance in the Medical field and manual counting of Red
Blood Cells in thin blood smear digital images.
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
National Institute of Health, N. "Red Blood Cells Count". 3-01-2014 [cited 2014
27-4-2014];
Available
from:
http://www.nlm.nih.gov/medlineplus/ency/article/003644.htm,1997.
Abbas, N. and Mohamad, D., "Automatic Color Nuclei Segmentation of
Leukocytes for Acute Leukemia". Research Journal of Applied Sciences,
Engineering and Technology, vol.7(14): pp. 2987-2993,2014.
Berge, H., Taylor, D., Krishnan, S., and Douglas, T.S. "Improved red blood cell
counting in thin blood smears". in Biomedical Imaging: From Nano to Macro,
2011 IEEE International Symposium on. IEEE,2011.
Buttarello, M. and Plebani, M., "Automated blood cell counts state of the art".
American Journal of Clinical Pathology, vol.130(1): pp. 104-116,2008.
OWAIS SHAIKH, M.G., NEHARIKA BHAT,ROSHAN SHETTY
"AUTOMATED RED BLOOD CELLS COUNT", in Department of Computer
Engineering, Rizvi College of EngineeringUniversity of Mumbai: Bandra(w),
Mumbai - 400050. pp. 29,2013-2014.
Nguyen, N.-T., Duong, A.-D., and Vu, H.-Q., "Cell splitting with high degree of
overlapping in peripheral blood smear". Int J Comp Theory Eng, vol.3(3),2011.
Mahmood, N.H., Lim, P.C., Mazalan, S.M., and Razak, M.A.A., "BLOOD
CELLS EXTRACTION USING COLOR BASED SEGMENTATION
TECHNIQUE". International Journal of Life Sciences Biotechnology and
Pharama Research, vol.2(2): pp. 233-240,2013.
grietinfo.in. "grietinfo.in".
[cited 2013 20-03-2013]; Available from:
grietinfo.in/projects/MAIN/BME2013/cd-8-project%20report_1_.pdf.3102,
Mahmood, N.H. and Mansor, M.A., "Red blood cells estimation using Hough
transform technique". Signal & Image Processing: An International Journal
(SIPIJ), vol.3(2): pp. 53-64,2012.
Ramesh, N., Salama, M.E., and Tasdizen, T. "Segmentation of haematopoeitic
cells in bone marrow using circle detection and splitting techniques". in
Biomedical Imaging (ISBI), 2012 9th IEEE International Symposium on.
IEEE,2012.
Buggenthin, F., Marr, C., Schwarzfischer, M., Hoppe, P.S., Hilsenbeck, O.,
Schroeder, T., and Theis, F.J., "An automatic method for robust and fast cell
detection in bright field images from high-throughput microscopy". BMC
bioinformatics, vol.14(1): pp. 297,2013.
Prasad, K., Winter, J., Bhat, U.M., Acharya, R.V., and Prabhu, G.K., "Image
analysis approach for development of a decision support system for detection of
malaria parasites in thin blood smear images". Journal of digital imaging,
vol.25(4): pp. 542-549,2012.
Naveed Abbas et. al. /IRICT (2014) 342-351
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
351
Wang, H., Zhang, H., and Ray, N. "Clump splitting via bottleneck detection". in
Image Processing (ICIP), 18th IEEE International Conference on. IEEE,2011.
Kumarasamy, S.K., Ong, S., and Tan, K.S., "Robust contour reconstruction of red
blood cells and parasites in the automated identification of the stages of malarial
infection". Machine Vision and Applications, vol.22(3): pp. 461-469,2011.
Wen, Q., Chang, H., and Parvin, B. "A Delaunay triangulation approach for
segmenting clumps of nuclei". in Biomedical Imaging: From Nano to Macro
ISBI'09. IEEE International Symposium on. IEEE,2009.
Makkapati, V.V. and Naik, S.K. "Clump splitting based on detection of dominant
points from contours". in Automation Science and Engineering, 2009. CASE 2009.
IEEE International Conference on. IEEE,2009.
Gurcan, M.N., Boucheron, L.E., Can, A., Madabhushi, A., Rajpoot, N.M., and
Yener, B., "Histopathological image analysis: A review". Biomedical Engineering,
IEEE Reviews in, vol.2: pp. 147-171,2009.
Cloppet, F. and Boucher, A. "Segmentation of overlapping/aggregating nuclei
cells in biological images". in Pattern Recognition, 2008. ICPR 2008. 19th
International Conference on. IEEE,2008.
Tulsani, H., "Segmentation using morphological watershed transformation for
counting blood cells". IJCAIT, vol.2(3): pp. 28-36,2013.
Ferro, L., Leal, P., Marques, M., Maciel, J., Oliveira, M.I., Barbosa, M.A., and
Quelhas, P., "Multinuclear Cell Analysis Using Laplacian of Gaussian and
Delaunay Graphs", in Pattern Recognition and Image AnalysisSpringer. pp. 441449,2013.
Hodneland, E., Kögel, T., Frei, D.M., Gerdes, H.-H., and Lundervold, A.,
"CellSegm-a MATLAB toolbox for high-throughput 3D cell segmentation".
Source code for biology and medicine, vol.8(1): pp. 1-24,2013.
Schmitt, O. and Reetz, S., "On the decomposition of cell clusters". Journal of
Mathematical Imaging and Vision, vol.33(1): pp. 85-103,2009.
Gonçalves, W.N. and Bruno, O.M., "Automatic system for counting cells with
elliptical shape". arXiv preprint arXiv:1201.3109,2012.
Kong, H., Gurcan, M., and Belkacem-Boussaid, K. "Splitting touching-cell
clusters on histopathological images". in Biomedical Imaging: From Nano to
Macro, 2011 IEEE International Symposium on. IEEE,2011.
Kong, H., Gurcan, M., and Belkacem-Boussaid, K., "Partitioning histopathological
images: an integrated framework for supervised color-texture segmentation and
cell splitting". Medical Imaging, IEEE Transactions on, vol.30(9): pp. 16611677,2011.
Köppen, M., Yoshida, K., and Valle, P.A. "Gestalt theory in image processing: A
discussion paper". in Proceedings.2007.
DPDx, "DPDx, Laboratory Identification of Parasites,Centers of diseases Control
and Prevention", D.a.m.b.C.s.D.o.P.D.a.M. (DPDM), Editor Web site developed
and maintained by CDC's Division of Parasitic Diseases and Malaria (DPDM):
1600 Clifton Rd. Atlanta, GA 30333, USA2002.
ABBAS, N. and MOHAMAD, D., "MICROSCOPIC RGB COLOR IMAGES
ENHANCEMENT FOR BLOOD CELLS SEGMENTATION IN YCbCr COLOR
SPACE FOR K-MEANS CLUSTERING". Journal of Theoretical & Applied
Information Technology, vol.54(1),2013.